Who supplies OpenAI's embodied data? A report points to two-year-old Maxinsights

A QbitAI investigation tries to answer where OpenAI's embodied AI training data comes from. OpenAI has not disclosed the supplier behind its egocentric datasets, the report says, but as Physical AI heats up one name keeps surfacing in industry discussion: Maxinsights, a Silicon Valley startup only two or three years old.

The scale figures in the piece are specific: more than 2 million hours of data delivered to customers, plus over 1.5 million hours carrying hand tracking and language annotation, with monthly collection capacity of 450,000 hours across a network spanning more than 2,000 real environments and 50,000 scenes on six continents.

The customer list is the part that draws attention. The report names Google DeepMind, Figure AI, 1X, Dyna Robotics and Genesis AI as partners, and says annual recurring revenue has sat at or near the $100 million mark for several consecutive months.

Two shipped models illustrate why the data matters. Dyna Robotics' Dyna-2, released in August, was pretrained on 1 million hours of human first-person video, and its prediction error on human motion fell about 12 percent as pretraining data scaled from 1,000 to 1 million hours; the report says Maxinsights was the largest supplier behind that corpus.

The second is Genesis AI's GENE-26.5 foundation model from May, which autonomously completed roughly 20 consecutive cooking steps in a kitchen, with many difficult skills learned from under an hour of task-specific robot data; the report says Maxinsights was its exclusive first-person data partner.

The economics are straightforward. Having robots collect their own manipulation data is expensive, slow and narrow in coverage, while head-mounted cameras recording everyday hand-object interaction are cheaper, more varied and independent of costly hardware, which is what makes million-hour datasets feasible.

The report also says the company expanded into the Chinese market in March, a sign that embodied data supply has moved past lab-scale collection into a cross-border business.

Caveats belong in the same breath: OpenAI has not confirmed the supplier, and the customer relationships come from public reporting and industry sourcing rather than company disclosure. The direction, though, is clear. In Physical AI, data collection networks are becoming a resource that has to be locked up in advance, much like compute.